Scholarlink

Abstract <p>This paper examines the architecture options of long short-term memory networks in the problem of forecasting financial time series using the example of cryptocurrency prices on minute, hourly, and daily ticks. The aim of the work was to study the influence of the activation function and hyperparameters of the long short-term memory network on forecast quality. Classic activation functions such as sigmoid, hyperbolic tangent, linear, rectified linear unit, as well as the nonclassical activation functions cloglog and loglog are considered.&#xa0;Hyperparameters define the architectural decisions of the neural network and affect the quality of the forecast and the speed of learning. This paper shows that the correct choice of the activation function and hyperparameters of long short-term memory networks is an important factor that influences the forecast quality and that the nonclassical activation functions cloglog and loglog should be used in long short-term memory network architectures for time series forecasting tasks along with classical activation functions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An Investigation of Activation Functions and Hyperparameters in Long Short-Term Memory Network Architectures for Financial Time Series Forecasting

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • Mikhail Ivanovich Kumskov,
  • Alina Rustiamovna Arslanova,
  • Anastasiia Gennadevna Makhova

摘要

Abstract

This paper examines the architecture options of long short-term memory networks in the problem of forecasting financial time series using the example of cryptocurrency prices on minute, hourly, and daily ticks. The aim of the work was to study the influence of the activation function and hyperparameters of the long short-term memory network on forecast quality. Classic activation functions such as sigmoid, hyperbolic tangent, linear, rectified linear unit, as well as the nonclassical activation functions cloglog and loglog are considered. Hyperparameters define the architectural decisions of the neural network and affect the quality of the forecast and the speed of learning. This paper shows that the correct choice of the activation function and hyperparameters of long short-term memory networks is an important factor that influences the forecast quality and that the nonclassical activation functions cloglog and loglog should be used in long short-term memory network architectures for time series forecasting tasks along with classical activation functions.